Skip to content

Burgers Viscous 1D

Download .py Back to chapter

This example solves the viscous Burgers equation with a manufactured exact solution and demonstrates RAD adaptive resampling — a technique that concentrates collocation points in high-residual regions to accelerate convergence.

Problem Setup

The PDE has the form u_t + u u_x = ν u_xx + f, with manufactured solution u_exact = e^{-t} sin(πx) and the corresponding forcing term f.

The domain [0,1] × [0,1] is discretised as a 2D rectangle with x (space) and t (time) treated as two spatial coordinates — the standard PINN formulation that makes full adaptive resampling available.

Step 1: Define the Resampling Strategy

RAD fires every 200 epochs (starting at epoch 500 once the network has formed a rough solution), replacing 20 % of working-set points with draws from the full 513-node mesh pool, biased toward regions of high PDE residual.

from jno import sampler

strategy = sampler.rad(
    resample_every=200,    # check for updates every 200 epochs
    resample_fraction=0.2, # replace 20 % of working-set points
    start_epoch=500,       # wait for a rough solution before resampling
    k=5,                   # cluster around the top-5 high-residual anchors
)

Step 2: Build the Space-Time Domain

mesh_size=0.05 on [0,1]² gives 513 interior nodes in the candidate pool. With sample=(60, None), the network trains on a 60-point working set — an 8× pool-to-sample ratio that gives RAD substantial room to move points into high-residual regions.

domain = 1 * jno.domain.rect(
    mesh_size=0.05,
    x_range=(0.0, 1.0),
    y_range=(0.0, 1.0),
)
vars_int = domain.variable("interior", sample=(60, None), resampling_strategy=strategy)
x, t = vars_int[0], vars_int[1]

# IC boundary: t = 0 (bottom face of the rectangle)
vars_bot = domain.variable("bottom", sample=(20, None))
x0, t0 = vars_bot[0], vars_bot[1]

Step 3: Define the Network and PDE

An MLP with 2 inputs (x, t) takes both coordinates directly. The x*(1-x) factor hard-encodes the Dirichlet BCs u(0,t) = u(1,t) = 0.

net = jno.nn.wrap(foundax.mlp(2, hidden_dims=64, num_layers=4, key=jax.random.PRNGKey(3)))
net.optimizer(optax.adam(optax.warmup_cosine_decay_schedule(init_value=0.0, peak_value=1e-3, warmup_steps=1, decay_steps=10, end_value=1e-5)))

u = net(x, t) * x * (1 - x)

u_t  = u.d(t)
u_x  = u.d(x)
u_xx = u_x.d(x)
pde  = u_t + u * u_x - ν * u_xx - source

Step 4: Train and Evaluate

u_0 = net(x0, t0) * x0 * (1 - x0)
ini = u_0 - jno.np.sin(π * x0)

crux    = jno.core([pde.mse, ini.mse])
history = crux.solve(5000)

_u, _u_exact = crux.eval([u, u_exact])
rel_l2 = float(jax.numpy.linalg.norm(_u - _u_exact) / (jax.numpy.linalg.norm(_u_exact) + 1e-8))

What To Notice

  • The nonlinear term u u_x creates residuals that are largest near x = 0.5 (the maximum of sin(πx)). After the warm-up phase, RAD progressively moves interior points toward that region.
  • The pool-to-sample ratio (8×) is what gives RAD room to relocate points. A ratio below 2× leaves too few candidates to draw from.
  • The initial-condition boundary (t = 0, sampled via "bottom") is kept fixed — only interior points are resampled.
  • For the classic inviscid-limit Burgers benchmark (ν = 0.01/π, no forcing, IC u(x,0) = -sin(πx)), the steep gradient forms near x = 0 at late times and RAD provides the largest gains. See adaptive/resampling.md for that setup.

Download full script Back to 04 Hyperbolic

Script Snippet

"""04 — 1-D viscous Burgers equation  (manufactured solution)"""

import foundax
import jax
import optax

import jno

π = jno.np.pi
ν = 0.05
T_end = 1.0

# ── Domain (1-D space × time) ─────────────────────────────────────────────────
domain = jno.domain.line(mesh_size=0.1, time=(0, T_end, 4))
x, t = domain.variable("interior")
x0, t0 = domain.variable("initial")

# ── Manufactured solution + source term ──────────────────────────────────────
u_exact = jno.np.exp(-t) * jno.np.sin(π * x)
source = jno.np.exp(-t) * (ν * π**2 - 1) * jno.np.sin(π * x) + (π / 2) * jno.np.exp(-2 * t) * jno.np.sin(2 * π * x)

# ── Network  (hard Dirichlet BCs via the x(1-x) factor) ──────────────────────
net = jno.nn.wrap(
    foundax.deeponet(
        n_sensors=1,
        coord_dim=1,
        n_outputs=1,
        n_layers=4,
        basis_functions=64,
        hidden_dim=48,
        key=jax.random.PRNGKey(3),
    )
)
net.optimizer(
    optax.adam(
        optax.warmup_cosine_decay_schedule(
            init_value=0.0, peak_value=1e-3, warmup_steps=10, decay_steps=5000, end_value=1e-5
        )
    )
)

u = net(t, x) * x * (1 - x)

# ── PDE residual:  u_t + u u_x − ν u_xx − f = 0 ──────────────────────────────
u_x = u.d(x)
pde = u.d(t) + u * u_x - ν * u_x.d(x) - source

# ── Initial condition ────────────────────────────────────────────────────────
u_0 = net(t0, x0) * x0 * (1 - x0)
ini = u_0 - jno.np.sin(π * x0)

# ── Solve ────────────────────────────────────────────────────────────────────
crux = jno.core([pde.mse, ini.mse])
crux.solve(5000)

_u, _u_exact = crux.eval([u, u_exact])
rel_l2 = float(jax.numpy.linalg.norm(_u - _u_exact) / (jax.numpy.linalg.norm(_u_exact) + 1e-8))
assert rel_l2 < 1e-1, f"relative L2 error too large: {rel_l2:.3e}"